Writing · Essays & notes
Long-form essays and shorter notes on technology, careers, books, history, psychology, faith, politics, money, learning, relationships and place. Written slowly, revised often.
- PsychologyEqual Trust Is Not a Strategy: What an AI Diplomacy Game Taught Me About Friends, Money and Sifuna's DilemmaMy AI agent trusted every player equally and lost most rounds, while another engineer's agent won every round. I have lost the same way in life, once with money a friend was meant to use to import a car for me. Here is the scoring arithmetic, the code, and what it says about Edwin Sifuna's choice for 2027.
- TechnologyBuild a Secure AI Agent Platform on Your Laptop: A Code WalkthroughMy last post explained how AWS secures AI agents on Amazon EKS. This one is the code. We rebuild every idea on a laptop with free tools, a model gateway, MCP tools behind per-persona rules, a sandbox for model-written code and end-to-end tracing, and test each one as we go.
- TechnologySecuring AI Agents on Amazon EKS: Seven Labs, One LessonI spent a day building an AI customer-service agent on Amazon EKS and growing it into one that writes code and opens pull requests. The agent barely changed. Everything that made it safe lived in the platform around it.
- PoliticsGovernment by Magic Spell: A Somali Story, a Kenyan Cabinet and the Race Rwanda WonIn a Somali short story we read for KCSE, a clan rules a country with a ten-year-old girl, a jinni and a spell that cures citizens of curiosity. The week Rwanda won Formula 1, Kenyans went looking for the spell in our own Cabinet.
- TechnologyAnatomy of a Full-Stack Web App: Following One Car From a Click to the Database and BackWhat actually happens when you click "Save" in a web app? Using a small car inventory app built with React, Spring Boot and MariaDB, we follow three everyday actions (listing, adding and deleting) from the button to the database and back, with diagrams and the real code.
- TechnologyWe Are All Living in Bird Box: KOX, a Newsmax Host and the Data We Can't SeeA Newsmax host compared American cities to Nairobi. Within hours Kenyans on X had his apology, and his phone number. A computer scientist's look at context collapse, doxxing and hacktivism, and how exposed every one of us already is.
- TechnologyWhat an AWS Platform Engineering Lab Taught Me About Running Apps in ProductionI joined AWS's online lab on Amazon EKS to see how applications should be deployed and managed in production. Seven lessons, from why even the build pipeline should go through Git to why a slow release should fail its canary, and why a platform makes AI safer.
- TechnologySpring Security Without WebSecurityConfigurerAdapter: Updating Spring Boot 2 Security Code for Spring Boot 4WebSecurityConfigurerAdapter no longer exists in Spring Security 7. Here is how a classic Spring Boot 2 security setup (in-memory users, database users and JWT login) looks on Spring Boot 4.1, side by side and tested.
- TechnologyYour Agent Has No Taste: 25 Design Resources for Vibe CodingAsk an AI agent for a modern landing page and you get the average of the internet. Taste is a deviation from the average. Here are 25 resources that give your agent something to deviate towards, checked, corrected and sorted into five jobs: see, say, build, move and finish.
- TechnologyWhat If Children Could See Maths Instead of Memorising It? How I Finally Got My Revenge on the RemainderWhen my son started to struggle with maths, I remembered how we dreaded "remainders" in Grade 2. So I built Hesabu Lab, a free 3D math game for kids where ten marbles become one crate and carrying, borrowing and remainders finally make sense.
- HistoryOne People of the Mara: Were the Abagusii, Abakuria and Their Neighbours Ever Separate Tribes?Kipsigis tradition calls Bureti and Sotik an "upper Kisii". Kuria elders say they lived with the Gusii at Gutura. The sources point to one people split by later labels, and the place names along the rivers of the Mau may show the road they took.
- TechnologyFinding Grape Varieties Without Labels: PCA and K-Means, Step by StepGiven the chemistry of 178 wines and no labels, can a machine discover that they come from three grape varieties? It can, but only if you get one preprocessing step right.
- TechnologyNaive Bayes From Scratch: Counting Words to Catch SpamNaive Bayes is a classifier you can compute with a pencil. We work one out by hand, see why smoothing matters, then build a tuned sentiment classifier in scikit-learn.
- TechnologySupport Vector Machines: Margins, Kernels and Why Scaling Comes FirstAn SVM draws the widest possible street between two classes. We build one on tumour measurements, see what unscaled features do to it, and learn what C and gamma control.
- TechnologyWhat Drives Medical Insurance Costs? Linear Regression, Read ProperlyWe fit a linear regression to 1,338 insurance bills, interpret every coefficient, check the assumptions, compare Lasso, and then let the residuals show us the model we should have built.
- TechnologyRandom Forests in Practice: More Trees, Misleading Importances and Rare ClassesRandom forests are the dependable workhorse of tabular ML. We test four things people get wrong about them, on real tumour data, and one of the answers surprised me.
- TechnologyDecision Trees, Explained Through a Diabetes ScreenA decision tree is a flowchart the computer writes for you. We build one on real clinical data, read its rules in plain English, and watch it overfit.
- TechnologyVibe Coding and the Return of the Software CrisisAI can now turn a prompt into working software. That is real progress, and it is also how the software crisis of the 1960s began: output outran understanding. Why engineering discipline matters more in the AI era, not less.